Documentation

A local-first time machine for AI-assisted coding. It watches your project, turns meaningful changes into checkpoints with real AI-written explanations, and gives you a genuine way back — a full or selective restore, or a fork to try something different — whenever an AI coding session goes somewhere you didn’t want.

It is not version control. It runs alongside git (or without it) and fills a different gap: why a change happened, tied to the actual prompt that caused it, plus a safety net for the moments in between commits — the exact window where a bad AI-driven edit is hardest to undo cleanly.

Every state whydid records — the initial snapshot, each checkpoint, every restore and fork — is a node in a real branching graph, parented on whatever came before it. A restore doesn’t discard your current work to get there: it safety-snapshots what’s on disk right now, automatically, before touching anything, so the state you were just in is always still there as its own node.

Files are stored content-addressed and deduplicated — only a full manifest is kept periodically, with everything else stored as deltas, reconstructed exactly on demand. A 10,000-checkpoint project doesn’t mean 10,000 copies of every unchanged file.

macOS, Linux, and Windows are all supported — not assumed to work the same way and left untested. Machine identification, OS keychain storage, and file-watching have each been checked against real platform differences rather than ported once and hoped for: Windows, for instance, names Claude Code’s session folders differently than macOS does, and locks open files in ways POSIX doesn’t — whydid accounts for both rather than assuming one platform’s behavior everywhere.

whydid requires Node.js 24+ (or 23.4+) for its built-innode:sqlitesupport. If your default node resolves to something older, the CLI detects this at startup and tells you exactly what’s installed and how to fix it, rather than failing with an unrelated error.

Coding agent support, honestly tiered:

• Claude Code — deepest integration. whydid reads Claude Code’s own local session transcripts directly: real prompts become real Sessions entries, checkpoints split exactly at prompt boundaries instead of waiting for a debounce timer, and a context-compaction event is detected and flagged (see Continuity in the dashboard).
• Codex and Cursor — hook-based. whydid init installs a small, standard hook config for each (.codex/hooks.json, .cursor/hooks.json) that reports session/prompt/stop events through each tool’s own supported hook interface — the same Sessions view, the same prompt-to-checkpoint linking, without reading any private file format.
• Any other AI coding tool, or plain manual edits — checkpointing, AI risk annotation, restore, fork, and diffs all work identically. whydid’s file watcher doesn’t care what made the change. What you won’t get without Claude Code, Codex, or Cursor specifically is prompt-level attribution — Sessions won’t show what you actually asked for, only what changed.
• AGENTS.md, CLAUDE.md, and Cursor’s own rules file are installed by whydid init regardless of which agent you use, so any tool that reads project-level instructions picks up whydid’s CONTEXT.md convention automatically.

Still stuck? Email hello@whydid.dev — a real response, not a ticket queue.

Your source, history, and snapshots stay on your machine — whydid’s own servers never receive your code. The only network calls whydid ever makes are to your own configured Anthropic or OpenAI account, using your own API key, and only when generating an AI annotation, chat answer, or build-in-public draft.

You control exactly what’s sent for those calls — your prompt, the diff, both, or neither — in Settings. Unchanged source and your full repository are never sent, regardless of these settings. whydid never logs or transmits your API key itself.